Textbook of AI

Aider is a command-line AI pair-programming tool launched by Paul Gauthier in May 2023. It connects to OpenAI, Anthropic, or local models, and lets the developer edit a real codebase by chatting with the LLM. Aider's distinctive engineering contribution is its careful study of edit formats, the wire format the model uses to express file modifications.

The workflow

$ aider main.py utils.py
aider> Refactor utils.py to use pathlib instead of os.path

Aider edits utils.py, runs tests, commits with message:
"refactor: replace os.path with pathlib in utils.py"

Aider:

1. Loads the repo into context.
2. Builds a repo map (a tree-sitter-derived summary of every file).
3. Sends the user request + relevant files + repo map to the LLM.
4. Parses the response in one of several edit formats.
5. Applies edits, runs lints/tests, and commits.

Edit formats

Gauthier's blog famously measures which edit format yields the highest SWE-bench-style success. Formats include:

Aider publishes a leaderboard of LLMs' code-editing skill on the same benchmark, regularly updated. This empirical rigour is highly unusual in the coding-agent space.

Distinctive features

• Git-native, every change is a commit; rollback is trivial.
• Auto-test loop, --auto-test runs your test command, feeds failures back to the model.
• Repo map with tree-sitter, summaries of all files via AST, so the model knows what's there even without loading every file.
• Lint integration, model is shown lint errors and asked to fix them.
• Chat mode vs apply mode, /architect mode separates planning from editing.

Modern relevance

Aider was the first open-source tool to demonstrate that LLMs could maintain a real codebase rather than write toy snippets. By 2025 it is the canonical citation in coding-agent research and shapes the design of OpenHands, Devin, OpenAI Codex CLI (2025), and Anthropic's Claude Code. Its leaderboard remains one of the most-trusted real-world evaluations of model coding ability.

Citation

Gauthier, P. (2023). Aider. https://aider.chat/.